Scalar inferences are commonly generated when a speaker uses a weaker expression rather than a stronger alternative, e.g., "John ate some of the apples" implies that he did not eat them all. This article describes a visual-world study investigating how and when perceivers compute these inferences. Participants followed spoken instructions containing the scalar quantifier "some" directing them to interact with one of several referential targets (e.g., "Click on the girl who has some of the balloons"). Participants fixated on the target compatible with the implicated meaning of "some" and avoided a competitor compatible with the literal meaning prior to a disambiguating noun. Further, convergence on the target was as fast for "some" as for the non-scalar quantifiers "none" and "all". These findings indicate that the scalar inference is computed immediately and is not delayed relative to the literal interpretation of "some". It is argued that previous demonstrations that scalar inferences increase processing time are not necessarily due to delays in generating the inference itself, but rather arise because integrating the interpretation of the inference with relevant information in the context may require additional time. With sufficient contextual support, processing delays disappear. (Contains 3 tables and 6 figures.)